Avi Goldfarb - The Economic Impact of AI - [Invest Like the Best, EP.321]

Thematic Section

1. AI as a Prediction Technology: Cost Declines Drive Application Explosion

Goldfarb argues that the essence of AI is prediction technology (in a statistical sense, using existing information to fill in missing information), and its exponentially declining costs will give rise to unprecedented application scenarios.

• Historical Analogy: Computers essentially only perform arithmetic. When the cost of arithmetic became low enough, problems that were not originally arithmetic (such as music and images) were redefined as arithmetic problems. Similarly, as a prediction technology, the declining cost of AI will redefine more problems (such as writing and diagnosis) as prediction problems.
• Data Support: Goldfarb points out that the core of ChatGPT is prediction—filling in missing information based on queries to generate coherent text. This was still hard to imagine five years ago when writing Prediction Machines.
• Key Judgment: "When prices drop by several orders of magnitude, we can identify extraordinary applications that were previously unimaginable." This means the impact of AI will far exceed current perceptions.

2. From "Point Solutions" to "System Solutions": Lessons from the Electricity Transition

Goldfarb emphasizes that the true value of AI lies not in simply replacing old processes (point solutions), but in redesigning entire systems (system solutions), which takes time—similar to the 40-year transition from steam power to electricity.

• Historical Context: Edison invented the light bulb in 1880, but it was not until the 1920s that half of U.S. factories and homes were electrified. Early factories merely replaced steam engines with electric motors (point solutions), saving only 5-15% in energy costs, which was not worth large-scale adoption. It was not until after 1900 that entrepreneurs realized electricity was a "distributed energy source" that could decouple the power source from machine location, allowing them to redesign factory layouts (system solutions) and unlock massive value.
• AI Analogy:
• Point Solution: For example, Derifin (a Canadian AI unicorn) uses machine learning to replace a bank's existing fraud prediction process, keeping the workflow unchanged but making it cheaper and better.
• Application Solution: For example, Ada Support uses AI to assist customer service representatives with standardized requests (such as password resets), embedding into the existing value chain but altering some processes.
• System Solution: For example, Uber (combining navigation prediction, digital dispatch, and demand forecasting to create a new transportation system) and digital advertising (shifting from fixed-rate cards to real-time bidding, giving rise to entirely new industry chains like DSPs and SSPs).
• Goldfarb's Judgment: "If you just replace an old process with a new one without changing the workflow, it may not be worth the effort. It is only worthwhile when you can do things in a completely different way and deliver new value." This means investors should focus on companies that are restructuring industry systems, rather than those merely doing "AI replacement."

3. From "Rules" to "Decisions": How Prediction Changes Organizational Logic

Goldfarb proposes that AI will shift organizations from being "rule-driven" (one-size-fits-all due to insufficient information) to "decision-driven" (flexible responses based on probabilistic judgments), but this requires coordinating a large number of supporting decisions.

• Mechanism Breakdown: In the absence of prediction, organizations rely on rules (e.g., "everyone must quarantine at home") to simplify management. Prediction technology provides probabilistic information (e.g., "you have a 36% chance of infection"), enabling organizations to make differentiated decisions (e.g., "only high-risk individuals quarantine"). However, decisions require supporting systems (e.g., sick pay, health data management, bio-waste disposal), otherwise they cannot be implemented.
• COVID Case: Goldfarb notes that COVID was, for most people, an "information problem" rather than a "health problem." If good prediction tools (such as rapid testing) had been available, full lockdowns could have been avoided. But even with such tools, companies would have needed to change compensation policies, privacy processes, etc., to shift from "rules" (shutting down) to "decisions" (only positive cases stay home).
• Key Judgment: "With prediction in place, we can actually conduct business as usual... but to move from rules to decisions, a large number of other decisions need to be coordinated both inside and outside the company." This means the implementation of AI is not just a technical issue but an organizational transformation issue.

4. Power Shifts and Disruption: Who Benefits, Who Loses

Goldfarb believes that AI will lead to a shift in power from traditional giants to new entrants, but data ownership and computing resources may concentrate power in the hands of a few companies.

• Disruption Paths:
• Demand-Side Disruption (Clayton Christensen style): New entrants start in low-end markets and gradually move upward. For example, AI enables low-skilled workers to perform high-skilled tasks (e.g., using ChatGPT for writing), potentially disrupting traditional professional services.
• Supply-Side Disruption (Rebecca Henderson style): Existing organizations are unable to adopt new technologies due to rigid processes. For example, Blockbuster did not fail to see the trend of digital distribution; rather, its franchise system could not support the transition.
• Beneficiaries:
• Infrastructure Providers: Such as NVIDIA (GPUs) and cloud service providers (AWS, Azure), benefiting from the explosion in AI computing demand.
• Holders of Complementary Assets: Such as pharmaceutical companies; if they own drugs for specific diseases, AI diagnostics can expand the patient pool and increase profits.
• Providers of Personalized Services: For example, in education, AI can enable tailored instruction, breaking the "grade-level" rule.
• Risk: Goldfarb warns: "If AI leads to a reduction in skill inequality in the future, but capital ownership is concentrated in the hands of a few, leading to massive inequality and abuse of power, that is what we should truly worry about." This means investors need to pay attention to systemic risks arising from data monopolies and computing power concentration.

5. Data Strategy: From "Collect Everything" to "Goal-Driven"

Goldfarb emphasizes that data is the core fuel for AI, but companies should avoid blindly collecting data and instead design data strategies backward from "prediction goals."

• Key Principle: "If you don't know what you want to predict, you will end up wasting millions of dollars creating an easy-to-use data interface that is utterly useless." Companies should first define the predictions needed for "system-level changes" and then trace back to what data is required.
• Data Acquisition Paths:

1. Purchase: Obtain from third parties.

2. Create: By launching point solutions or application solutions, collect data while serving customers. For example, Tesla collected driving data by selling cars with sensors before achieving full autonomous driving.

3. Simulate: Use "digital twins" or reinforcement learning to generate data. For example, Singapore uses a digital simulation of the city to assess the impact of new buildings on traffic.

• Goldfarb's Judgment: "Scale and quality both matter, but both must serve a specific purpose." This means investors should focus on companies with a clear data strategy, rather than those that simply possess massive amounts of data.

---

Mentioned Positions

Position	Guest Stance	Key Data
Derifin	Bullish (point solution success story)	Canada's first AI unicorn, replacing banks' legacy fraud prediction processes with machine learning
Ada Support	Bullish (application solution case)	Helped Zoom handle a >10x surge in customer service inquiries after March 2020
Uber/Lyft	Bullish (system solution case)	Combined navigation prediction, digital dispatch, and demand forecasting to create a new transportation system
NVIDIA	Bullish (infrastructure beneficiary)	All cloud service providers run NVIDIA GPUs, with computing demand exploding
Tesla	Bullish (data strategy case)	Collects driving data by selling sensor-equipped vehicles, preparing for full autonomous driving
Blockbuster	Risk warning (supply-side disruption case)	Failed to transition to digital distribution due to an uncoordinated franchise system
Zoom	Neutral (mentioned as Ada customer)	Customer service inquiries surged >10x after March 2020

---

Judgments Worth Remembering

1. AI is a prediction technology, and its declining cost will redefine problems (Goldfarb): When the cost of prediction becomes low enough, problems that were not originally prediction-based (e.g., writing, diagnosis) will be redefined as prediction problems. ChatGPT proves that writing is essentially a prediction problem—predicting the next word based on a query.

2. The true value of AI lies in system-level transformation, not simple substitution (Goldfarb): The electricity transition took 40 years because early efforts merely replaced steam engines with electric motors (point solutions), and value was only unlocked when factories were redesigned (system solutions). The same applies to AI; Uber and digital advertising are examples of system solutions.

3. The shift from "rules" to "decisions" requires coordinating a large number of supporting decisions (Goldfarb): The COVID case shows that even with rapid testing (a prediction tool), companies needed to change processes related to compensation, privacy, and waste disposal to move from "shutting down" to "only positive cases staying home." AI deployment is an organizational change issue.

4. AI will disrupt traditional professional services but may create more jobs (Goldfarb): ChatGPT enables non-professionals to complete writing tasks, similar to how GPS allowed non-professional drivers to provide taxi services. This could result in "thousands harmed, millions benefited."

5. Data strategy should start from prediction goals, not blind collection (Goldfarb): Tesla collects driving data by selling cars equipped with sensors, preparing for full autonomous driving. Companies should first define "what to predict" and then design the data acquisition path (purchase, creation, simulation).

6. AI may exacerbate capital ownership concentration, leading to new forms of inequality (Goldfarb): If AI infrastructure (computing power, data) is monopolized by a few companies, skill inequality may decrease, but capital inequality will worsen. This is the most concerning risk.

7. Industries that "compensate for customer failures" are most vulnerable to AI disruption (Goldfarb): Airport shopping, dining, and other amenities essentially compensate for passengers' waiting time. If AI reduces waiting (e.g., more efficient security checks, predictive maintenance), these compensatory services will disappear. Investors should look for such "failure compensation" industries.

8. AI shifts decision-making from binary (yes/no) to decimal (probability assessment) (Goldfarb): Prediction machines output "36% probability" instead of "yes or no," forcing organizations to confront probabilities and biases. This is a natural advantage for investors but a new challenge for most industries.